Attributing water shortage in the western U.S. with explainable AI emulators of process-based hydrologic models
Water shortages across the western United States are driven both by natural factors, such as temperature and precipitation, and by human factors, such as population density, irrigation efficiency, cropland, and water-use intensity. This project uses the University of New Hampshire Water Balance Model (WBM) to quantify how each of these drivers contributes to water shortage. Because full sensitivity analyses of a process-based model are computationally impractical, I developed a deep learning emulator that captures WBM behavior with high fidelity and far greater efficiency, and then applied multiple explainable artificial intelligence (XAI) methods to quantify each input’s contribution. Different XAI methods can disagree even for the same model and data, so I combine their results into consensus-strength maps that show where the methods agree on the main drivers, and quantify how unstable and uncertain these attributions are.
- Developed a deep learning emulator of the WBM for the western U.S.
- Applied and compared multiple XAI methods to attribute water shortage to climate and socioeconomic drivers
- Combined explanations into consensus-strength maps to communicate where attributions are robust and where they are uncertain
Blog post: One Model, Many Stories: Why Your AI’s Explanation Might Be Misleading You
Research spotlight: People of PCHES